Allow custom priors and likelihood in DelayedSaturated MMM

[cetagostini]

Hey guys, I'm applying the same configuration of the models on CLV to the MMM models, the idea is to be able to pass priors and custom distributions based on the user needs.

Everything was working properly on the mmm example notebook, except for the fit() method 😅 Here is the traceback, looks like is not directly related to the changes:

---------------------------------------------------------------------------
NotImplementedError                       Traceback (most recent call last)
<ipython-input-26-e8614295d821> in <cell line: 1>()
----> 1 mmm.fit(X=X, y=y, target_accept=0.95, chains=4, random_seed=rng)

10 frames
/content/pymc-marketing/pymc_marketing/model_builder.py in fit(self, X, y, progressbar, predictor_names, random_seed, **kwargs)
    475             with self.model:
    476                 sampler_args = {**self.sampler_config, **kwargs}
--> 477                 self.idata = pm.sample(**sampler_args)
    478 
    479         X_df = pd.DataFrame(X, columns=X.columns)

/usr/local/lib/python3.10/dist-packages/pymc/sampling/mcmc.py in sample(draws, tune, chains, cores, random_seed, progressbar, step, nuts_sampler, initvals, init, jitter_max_retries, n_init, trace, discard_tuned_samples, compute_convergence_checks, keep_warning_stat, return_inferencedata, idata_kwargs, nuts_sampler_kwargs, callback, mp_ctx, model, **kwargs)
    649 
    650     initial_points = None
--> 651     step = assign_step_methods(model, step, methods=pm.STEP_METHODS, step_kwargs=kwargs)
    652 
    653     if nuts_sampler != "pymc":

/usr/local/lib/python3.10/dist-packages/pymc/sampling/mcmc.py in assign_step_methods(model, step, methods, step_kwargs)
    209     methods_list: List[Type[BlockedStep]] = list(methods or pm.STEP_METHODS)
    210     selected_steps: Dict[Type[BlockedStep], List] = {}
--> 211     model_logp = model.logp()
    212 
    213     for var in model.value_vars:

/usr/local/lib/python3.10/dist-packages/pymc/model/core.py in logp(self, vars, jacobian, sum)
    738             return logp_factors
    739 
--> 740         logp_scalar = pt.sum([pt.sum(factor) for factor in logp_factors])
    741         logp_scalar_name = "__logp" if jacobian else "__logp_nojac"
    742         if self.name:

/usr/local/lib/python3.10/dist-packages/pymc/model/core.py in <listcomp>(.0)
    738             return logp_factors
    739 
--> 740         logp_scalar = pt.sum([pt.sum(factor) for factor in logp_factors])
    741         logp_scalar_name = "__logp" if jacobian else "__logp_nojac"
    742         if self.name:

/usr/local/lib/python3.10/dist-packages/pytensor/tensor/math.py in sum(input, axis, dtype, keepdims, acc_dtype)
   2502     """
   2503 
-> 2504     out = Sum(axis=axis, dtype=dtype, acc_dtype=acc_dtype)(input)
   2505 
   2506     if keepdims:

/usr/local/lib/python3.10/dist-packages/pytensor/graph/op.py in __call__(self, *inputs, **kwargs)
    302         """
    303         return_list = kwargs.pop("return_list", False)
--> 304         node = self.make_node(*inputs, **kwargs)
    305 
    306         if config.compute_test_value != "off":

/usr/local/lib/python3.10/dist-packages/pytensor/tensor/elemwise.py in make_node(self, input)
   1384 
   1385     def make_node(self, input):
-> 1386         input = as_tensor_variable(input)
   1387         inp_dims = input.type.ndim
   1388         inp_dtype = input.type.dtype

/usr/local/lib/python3.10/dist-packages/pytensor/tensor/__init__.py in as_tensor_variable(x, name, ndim, **kwargs)
     47 
     48     """
---> 49     return _as_tensor_variable(x, name, ndim, **kwargs)
     50 
     51 

/usr/lib/python3.10/functools.py in wrapper(*args, **kw)
    887                             '1 positional argument')
    888 
--> 889         return dispatch(args[0].__class__)(*args, **kw)
    890 
    891     funcname = getattr(func, '__name__', 'singledispatch function')

/usr/local/lib/python3.10/dist-packages/pytensor/tensor/__init__.py in _as_tensor_variable(x, name, ndim, **kwargs)
     54     x: TensorLike, name: Optional[str], ndim: Optional[int], **kwargs
     55 ) -> "TensorVariable":
---> 56     raise NotImplementedError(f"Cannot convert {x!r} to a tensor variable.")
     57 
     58 

NotImplementedError: Cannot convert None to a tensor variable.

I'll be debugging during the week but if anyone has an idea about what could be, perhaps I can save some time 👍

---

📚 Documentation preview 📚: https://pymc-marketing--397.org.readthedocs.build/en/397/

[ricardoV94]

That's a weird one. Can you do pytensor.dprint(...) of the model build by the MMM class?

[ricardoV94]

@cetagostini Is there a test I can run to investigate?

[cetagostini]

Hey guys! @juanitorduz helped me with the test, I think everything seems to work now.

@ricardoV94 Let me know! 🚀

[ricardoV94]

Small suggestions, big picture looks great 👍

[cetagostini]

@ricardoV94 Hey I tried your recommended code based on a new test for the config and it didn't work.

[juanitorduz]

We are so close @cetagostini 🔥 !

[juanitorduz]

It seems the regex match is still failing

AssertionError: Regex pattern "The distribution used for the likelihood is not allowed. Please, use one of the following distributions: ['Normal', 'StudentT', 'Laplace', 'Logistic', 'LogNormal', 'Wald', 'TruncatedNormal', 'Gamma', 'AsymmetricLaplace', 'VonMises']." does not match "The distribution used for the likelihood is not allowed. Please, use one of the following distributions: ['Normal', 'StudentT', 'Laplace', 'Logistic', 'LogNormal', 'Wald', 'TruncatedNormal', 'Gamma', 'AsymmetricLaplace', 'VonMises'].". Did you mean to re.escape() the regex?

I guess you need to escape the [ and ] characters 😅

[cetagostini]

It seems the regex match is still failing

AssertionError: Regex pattern "The distribution used for the likelihood is not allowed. Please, use one of the following distributions: ['Normal', 'StudentT', 'Laplace', 'Logistic', 'LogNormal', 'Wald', 'TruncatedNormal', 'Gamma', 'AsymmetricLaplace', 'VonMises']." does not match "The distribution used for the likelihood is not allowed. Please, use one of the following distributions: ['Normal', 'StudentT', 'Laplace', 'Logistic', 'LogNormal', 'Wald', 'TruncatedNormal', 'Gamma', 'AsymmetricLaplace', 'VonMises'].". Did you mean to re.escape() the regex?

I guess you need to escape the [ and ] characters 😅

Yes, the string needed to be corrected. Now it's solved!

[cetagostini]

Hey there, team!

I just wanted to give a quick shoutout to everyone who contributed to this PR. It truly was a group effort and a testament to why the open-source community it's so great and moves so quickly.

It all started with an idea, which I drew inspiration from @nialloulton code. @MarkusSagen also provided some valuable insights, as he was working on something similar. @ricardoV94 supervised each step we took, ensuring that our code was not only readable but also bulletproof.

Finally after struggling with the tests, and a crazy chatgpt who was throwing out crazy nonsense (Instead of help), Mr. @juanitorduz stepped up and helped us cross the finish line with the missing unit tests.

Tremendous teamwork and support, everyone covering our backs in each area, moving the PR. I'm adding everyone as co-authors, because this PR deserves it.

Let's keep up the fantastic work!

[ricardoV94]

Great work @cetagostini

[juanitorduz]

"This branch cannot be rebased due to conflicts" ? Which conflicts?

[ricardoV94]

"This branch cannot be rebased due to conflicts" ? Which conflicts?

There were some merge pull commits from the main branch (instead of merge rebase)

[ricardoV94]

It's still possible to squash though (just did)